1. Designing Constraint-Based Agents Alan Mackworth

2. Four Perspectives on Designing Constraint-Based Agents The evolution of the pivotal role of constraints in intelligent systems: from static to dynamic constraints A theory of constraint-based agent design and a corresponding experiment in robot architecture The design of two assistive technology prototypes for people with physical and mental disabilities who are living with significant additional constraints Our collective failure to recognize and satisfy various constraints explains why many of the worlds we live in are unsustainable - out of kilter.

3. The Dynamics of Evolution Agents Models of Agents AI Constraint Satisfaction Thesis: Constraint satisfaction is central to intelligent behavior.

4. Constraint Satisfaction Problems CSP = UBC AIspace CSP Solver: www.AIspace.orgwww.AIspace.org

5. Sudoku Puzzle as a CSP Constraints: Each row, column and 3x3 group is a permutation of {1,2, …,9}.

6. Sudoku Puzzle as a CSP

7. Constraint Satisfaction Problems Network Consistency Algorithms Learn local inconsistencies – use them to prune search space efficiently: Arc consistency Path consistency k-consistency … (Fikes, Waltz, Montanari, Mackworth, Freuder, …) Constraint Programming New field for logistics, planning, scheduling, combinatorial optimization, ….

8. Pure Good Old Fashioned AI and Robotics (GOFAIR) Meta-assumptions Single agent Executes actions serially Deterministic world Fully observable, closed world Agent has perfect internal model of infallible actions and world dynamics Perception needed only to determine initial world state Perfect plan to achieve goal obtained by reasoning, and executed open loop

9. CSPs and GOFAIR CSPs as simple exemplar of GOFAIR. In pure GOFAIR, perfect model of the world, and its dynamics, in agent’s head. The agent is an Omniscient Fortune Teller. The agent’s world model and the world in perfect correspondence. We confused the agent’s world model and the world.

10. A Robot in the World

11. Classic Horizontal Architecture Good Old Fashioned Artificial Intelligence and Robotics (GOFAIR)

12. The Demise of GOFAIR GOFAIR robots succeed in blocks worlds and factories but can’t play soccer! Don’t let them into your home without adult supervision. “Life is what happens to you when you are busy making other plans.” (Lennon, 1980) An intelligent robot must be both proactive and responsive. Proactive: Acts to execute short-term and long-term plans and achieve goals in priority order, … Responsive: Reacts in real-time to changes in the environment, threats to safety, other agents, …

13. Beyond GOFAIR to Soccer Robot soccer, as a domain, breaks all the meta-assumptions of GOFAIR.

14. UBC Robot Soccer Dynamos (1992)

15. UBC Robot Soccer Dynamites (1992)

16. The Dynamites: Two on Two Soccer Playing Robots (UBC, 1993) Zeno & Heraclitus vs. Hegel & Leibnitz (Monty Python)

17. RoboCup (Kitano et al., Stone & Veloso, …)

18. From Sudoku to Soccer and Beyond SudokuSoccer Number of agents123 CompetitionNoYes CollaborationNoYes Real timeNoYes DynamicsMinimalYes ChanceNoYes OnlineNoYes Planning HorizonsNoYes Situated PerceptionNoYes Partially ObservableNoYes Open WorldNoYes LearningSomeYes

19. From GOFAIR to Situated Agents To build a proactive, responsive agent/robot we cannot just glue a proactive, modified GOFAIR planner on top of a reactive, control- theoretic controller. (But that is how we built the first robot soccer controllers.) Abandon the meta-assumptions of GOFAIR but keep constraint satisfaction. Constraints are now dynamic, coupling the agent and its environment e.g. kicking a soccer ball Constraints are the key to a uniform architecture. We need a new theory of constraint-based agents.

20. Robot Friends Stefan Eissing

21. RoboCars: DARPA Urban Challenge “Junior” “Boss” ( Stanford Racing Team, 2007) ( CMU-GM Tartan Racing Team, 2007)

22. Can We Trust Robots? Can they do the right thing? Will they do the right thing? Will they be autonomous, with free will, intelligence and consciousness? Do we need robot ethics, for us and for them?

23. What We Need Any ethical discussion presupposes we (and robots) can: Model robot structure and functionality Predict consequences of robot commands and actions Impose requirements on robot actions such as goal reachability, safety and liveness (absence of deadlock and livelock) Determine if a robot satisfies those requirements (almost always)

24. Theory Wanted We need a theory with: 1.Language to express robot structure and dynamics 2.Language for constraint-based specifications 3.Method to determine if a robot will (be likely to) satisfy its specifications, connecting 1 to 2

25. Constraints on an Agent To thrive, an agent must satisfy dynamic constraints deriving from four sources: A.Its internal structure B.Its goals and preferences C.Its external environment D.The coupling between its internal and external worlds The life of any agent who does not respect and satisfy those constraints will be out of balance.

26. A Robot Agent in the World

27. A Constraint-Based Agent Situated Agents Constraint Satisfaction Prioritized Constraints e.g. Asimov’s laws CBA Structure Constraint Solver

28. Vertical Architecture (Albus, Brooks,…)

29. Dynamic Constraint Satisfaction We say the coupled agent-environment system satisfies a constraint if the constraint’s solution set, in the phase space of the coupled hybrid dynamical system, is an attractor of the system, as it evolves. DCS subsumes the CSP model.

30. Formal Methods The CBA framework consists of: 1.Constraint Net (CN) → system modelling 2.Timed -automata → behaviour specification 3.Model-checking and Liapunov methods → behaviour verification A (Zhang & Mackworth, 1993; …, 2003)

31. A CN Program CN = Example: CN of a differential equation

32. Timed -automata Behaviour is the set of traces Trace is a sequence: Time → Value Domain Set of sequences defines a language If behaviour is a language then its specification is an automaton that accepts that language A Trace: x = sin(t) Constraint G: x ≥0 Example-1: Example-2:

33. robot that traverses a maze (Zhang and Mackworth, 1994) two handed robot that assembles objects (Zhang and Mackworth, 1994) elevator simulation (Ying Zhang and Mackworth, 1999) soccer-playing robots (Yu Zhang and Mackworth, 1998) Prior Work on CBA Framework Modelled in CNJ (Song and Mackworth, 2002) – (Montgomery and Mackworth, 2003)

34. Case Study A situated robot, Ainia, has the task of repeatedly finding, tracking, chasing and kicking a ball. (Muyan-Özçelik & Mackworth, 2004)

35. Prioritized Constraints: Ball-In-Image → I Ball-In-Center → C Base-Heading-Pan → H Robot-At-Ball → A I > C > H > A Ainia and Prioritized Constraints Aina named after an Amazon warrior. “Ainia” means “swiftness”

36. CBA in CN ← CBA Structure ↑ Control Synthesis with Prioritized Constraints Constraint1 > Constraint2 > Constraint3 >

37. Prioritized Constraints in CBA The Constraint-Based Agent (CBA) framework with prioritized constraints, using the Constraint Nets in Java (CNJ) tool, is an effective methodology for modeling and building situated agents in the real world. Using prioritized constraints we can get reliable situated, sequenced, reactive visuomotor behaviour

38. The CNJ Tool

39. Behaviour Verification Checks whether: Behaviour Specification For Ainia we use a formal model-checking method with Liapunov functions

40. Includes following modules: Tilt, BaseTranslation, Camera, BaseBump, Pan, BaseRotation. CN of Ainia modelled in CNJ Includes following modules: Ball, Kicker, Wall, Obstacle.

41. Only keep the Controller module Environment module → real world Body module → physical robot body Software of Physical Ainia median filter → connected components → measure P 2 /A Reports Commands Camera Views External View

42. Findings The robot always eventually kicks the ball repeatedly, both in simulation and experimentally. We can prove that the robot always eventually kicks the ball repeatedly. The Constraint-Based Agent approach with prioritized constraints is an effective framework for robot controller construction for a simple task. It is a solution to “The Problem of the Serial Ordering of Behavior” (Lashley, 1951).

43. Uncertainty? Where? Uncertainty in the Dynamics Measurement Noise TARGET Environmental Disturbances Control Force Probabilistic Constraint Nets (St-Aubin & Mackworth, 2004)

44. Observations Simple idea: constraint satisfaction is a way to achieve intelligent, proactive and reactive behaviour. The formal prioritized constraint framework allows for the emergence of robust, goal-seeking behaviours. Contributes to a foundation of frameworks for robot ethics. Yes, robots can do the right thing (sometimes).

45. Wayfinding for Wheelchair Users (Yang & Mackworth, 2007)

46. Wayfinding in a Building (Yang and Mackworth, 2007)

47. Satisfying Mobility Constraints in NOAH Navigation and obstacle avoidance in a collaboratively controlled smart wheelchair (Viswanathan et al., 2008)

48. Sustainability Sustainability The ability to maintain balance of a process in a system Ecological Sustainability The ability of an ecosystem to maintain ecological processes, functions, biodiversity and productivity into the future Human Sustainability The ability to meet the needs of the present without compromising the ability of future generations to meet their own needs

49. Constraint-based Sustainability Sustainable systems must satisfy physical, chemical, biological, psychological, economic, and social constraints. Consider constraints such as energy supply, waste management, GHG, ocean acidity, climate, ecological footprint, biodiversity, habitat, harvesting and equity. Sustainability = Constraint Satisfaction

50. Planetary Boundaries as Constraints [Source: “Planetary Boundaries: Exploring the Safe Operating Space for Humanity”, Rockström et al., Ecology and Society (2009)]

51. Safe Operating Space Simple conceptual model: 10-dimensional phase space for the planetary dynamical system. The safe operating space is defined by an envelope within which those biophysical constraints (x < x limit ) are all satisfied. Many of the subsystems display resilience in that they resist movement away from homeostasis, moving back to an equilibrium - Gaia principle (Lovelock). But disturbance beyond the resilient limit can tip the system into a new basin of attraction which may represent abrupt environmental change e.g. collapse.

52. How Can We Help? AI, CS and computation in service to people and our planet Global crises of poverty and inequity, overdevelopment, climate change, health, education, aging, … Teaching and learning tools for CS & AI: games, … Sustainability: smart cars, smart traffic control, scenario modeling, auctions for carbon credits, smart energy controllers, distributed sensors, smart communication systems, teleconferencing, … Assistive technology: assisted perception, cognition and action for seniors and disabled, smart wheelchairs, companions, nurse assistants, …

53. Constraints Shall Make You Free Every task involves constraint, Solve the thing without complaint; There are magic links and chains Forged to loose our rigid brains. Structures, strictures, though they bind, Strangely liberate the mind. — James Falen

54. Thanks Thanks to Rod Barman, Le Chang, Johan de Kleer, Pooyan Fazli, Gene Freuder, Bill Havens, Kletnathee Imhiran, Stewart Kingdon, Jim Little, Valerie McRae, Jefferson Montgomery, Pinar Muyan- Özçelik, Dinesh Pai, Fengguang Song, Michael Sahota, Ray Reiter, Robert St-Aubin, Pooja Viswanathan, Suling Yang, Ying Zhang and Yu Zhang. Funding: Canada Research Chair, IRIS NCE, Precarn, PWIAS, CIFAR, NSERC